Remove no-longer used type argument

example/36_permute/permute_fp16.cpp

@@ -6,8 +6,8 @@
 using ADataType = F16;
 using BDataType = F16;
 
-using DevicePermuteInstance = ck::tensor_operation::device::
-    DevicePermute<ADataType, BDataType, PassThrough, 4, 8, S<8>, S<1>>;
+using DevicePermuteInstance =
+    ck::tensor_operation::device::DevicePermute<ADataType, BDataType, PassThrough, 4, 8, 8, 1>;
 
 #include "run_permute_example.inc"
 

include/ck/tensor_operation/gpu/device/device_permute.hpp

@@ -73,23 +73,16 @@ template <typename InDataType,
           typename ElementwiseOperation,
           index_t NumDim,
           index_t MPerThread,
-          typename InScalarPerVectorSeq,
-          typename OutScalarPerVectorSeq>
+          index_t InScalarPerVector,
+          index_t OutScalarPerVector>
 struct DevicePermute : detail::DevicePermuteBase<DevicePermute<InDataType,
                                                                OutDataType,
                                                                ElementwiseOperation,
                                                                NumDim,
                                                                MPerThread,
-                                                               InScalarPerVectorSeq,
-                                                               OutScalarPerVectorSeq>>
+                                                               InScalarPerVector,
+                                                               OutScalarPerVector>>
 {
-    static constexpr int NumInput  = 1;
-    static constexpr int NumOutput = 1;
-
-    static_assert(NumInput == InScalarPerVectorSeq::Size() &&
-                      NumOutput == OutScalarPerVectorSeq::Size(),
-                  "Tuple size is inconsistent with the number of in/out!");
-
     using InDataTypePointer  = const InDataType*;
     using OutDataTypePointer = OutDataType*;
 
@@ -156,8 +149,8 @@ struct DevicePermute : detail::DevicePermuteBase<DevicePermute<InDataType,
                                             OutDataTypePointer,
                                             ElementwiseOperation,
                                             MPerThread,
-                                            InScalarPerVectorSeq,
-                                            OutScalarPerVectorSeq>;
+                                            InScalarPerVector,
+                                            OutScalarPerVector>;
 
     struct Argument : public BaseArgument
     {
@@ -243,12 +236,12 @@ struct DevicePermute : detail::DevicePermuteBase<DevicePermute<InDataType,
         };
 
         bool valid = true;
-        if(!IsScalarPerVectorValid(arg.inLengths_, arg.inStrides_, InScalarPerVectorSeq::At(0)))
+        if(!IsScalarPerVectorValid(arg.inLengths_, arg.inStrides_, InScalarPerVector))
         {
             valid = false;
         }
 
-        if(!IsScalarPerVectorValid(arg.inLengths_, arg.outStrides_, OutScalarPerVectorSeq::At(0)))
+        if(!IsScalarPerVectorValid(arg.inLengths_, arg.outStrides_, OutScalarPerVector))
         {
             valid = false;
         }

include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp

@@ -32,17 +32,10 @@ template <typename InGrid1dDesc,
           typename OutDataTypePointer,
           typename ElementwiseOperation,
           index_t MPerThread,
-          typename InScalarPerVectorSeq,
-          typename OutScalarPerVectorSeq>
+          index_t InScalarPerVector,
+          index_t OutScalarPerVector>
 struct GridwisePermute
 {
-    static constexpr index_t NumInput  = 1;
-    static constexpr index_t NumOutput = 1;
-
-    static_assert(NumInput == InScalarPerVectorSeq::Size() &&
-                      NumOutput == OutScalarPerVectorSeq::Size(),
-                  "Tuple size is inconsistent with the number of in/out!");
-
     static constexpr auto I0 = Number<0>{};
 
     static constexpr auto thread_buffer_desc_m =
@@ -83,11 +76,11 @@ struct GridwisePermute
                                              InDataType,
                                              decltype(in_grid_1d_desc),
                                              decltype(thread_buffer_desc_m),
-                                             Sequence<MPerThread>,        // SliceLengths
-                                             Sequence<0>,                 // DimAccessOrder
-                                             0,                           // SrcVectorDim
-                                             InScalarPerVectorSeq::At(0), // ScalarPerVector
-                                             1,                           // SrcScalarStrideInVector
+                                             Sequence<MPerThread>, // SliceLengths
+                                             Sequence<0>,          // DimAccessOrder
+                                             0,                    // SrcVectorDim
+                                             InScalarPerVector,    // ScalarPerVector
+                                             1,                    // SrcScalarStrideInVector
                                              false>{in_grid_1d_desc, thread_global_offset};
 
         auto out_global_store =
@@ -99,7 +92,7 @@ struct GridwisePermute
                                                Sequence<MPerThread>, // SliceLengths
                                                Sequence<0>,          // DimAccessOrder
                                                0,                    // SrcVectorDim
-                                               OutScalarPerVectorSeq::At(0),
+                                               OutScalarPerVector,
                                                InMemoryDataOperationEnum::Set,
                                                1,
                                                false>(